The long-term aim of this project is to fold proteins "in silico". The proposed research involves a variety of different theoretical and computational methods as well as a collaboration with an experimentalist with whom we have a long- standing relationship. The five specific aims are to be achieved by answering these questions: 1. How accurately do all-atom simulations reproduce the hydrophobic interactions? Encouraged by the abilityof molecular dynamics with simple all-atom energy functions to reproduce the hydrophobic interaction quantitatively, we will investigate the effect of temperature and pressure on our calculations. We will also examine the time- averaged arrangement of water molecules around a benzene molecule. 2. How do short chains of connected hydrophobic amino acids collapse? Having reproduced the clustering behavior of small hydrophobic molecules in molecular dynamics simulations, we plan to examine the role of the polypeptide chain in hydrophobic collapse by extending our techniques to short artificial sequences of the form GGGhGhGGGhhhGhhGGGhGG (h is a hydrophobic amino acid). 3. Do a-helix unfolding simulations agree in detail with experiment? Working with our experimental collaborator, Dr. William Eaton, we plan to study the rate of helix unfolding. This will involve new experimental work as well as massive amounts of computer time. With the help of Dr. Vijay Pande, we will deploy Encad as a massively parallel simulation system via thefolding@home client-server scheme. 4. Can simulated annealing minimization with special potentials generate better decoys? With our greater understanding of the need to develop a potential that captures the free-energy of a protein, we plan to develop terms for solvation and cooperativity. Use of a fast simulated annealing code will enable different potentials to be prototyped and tested rapidly. 5. Can decoy discrimination be improved by better clustering and the use of sequence homologues? Discrimination of near-native folds from less good structures has become increasingly important with our expanding ability to generate native-like decoy sets. We propose new methods involving homologous sequences, non-linear mapping of conformation space and robust clustering.